Knowledge baseSourcesDavid Konitzny (Peec AI)

How unique are domains & URLs in ChatGPT query fan-outs?

Author David Konitzny (Peec AI) Date 2026-08-24 Model state GPT-5.6, 500 brand-free prompts daily over 17 days, as of 24 Aug 2026 Open source →

Key findings

Konitzny runs 500 brand-free prompts daily for 17 days under GPT-5.6 and is the first to systematically separate the domain from the URL level of fan-out restrictions. Result: domain restrictions dominate (91.9%) and recur - a hard core of 18.1% of domains returns near-daily. URL paths are rare (9.1%) and almost always one-offs. His reading: the domain choice is learned and stable in the model, the specific page is re-decided per run.

Why this source matters for the project

The persistence analysis delivers a measurement idea our tracker can serve daily: domain recurrence per question as its own metric. It also supports the thesis behind Share of fan-outs with domain restriction that the domain choice is made before retrieval - whoever is absent there is never searched in the first place.

Checked against our data

Our five measurement days show the same picture in miniature: 86 domain parameters, all without a path; 25 unique domains, of which 28% are one-offs and a hard core of 9 domains (36%) appears on 4-5 of 5 days - openai.com, reuters.com and tagesschau.de on all five. Our core is proportionally larger than his (36% vs 18.1%), plausibly so: 26 frozen questions instead of 500 rotating prompts - the domain choice repeats per question. The core domains are established sources throughout. We do not see the 9.1% path share - conceivably an effect of the German prompt set or the sample size; noted as an open check.

ClaimStatusEvidence
91.9% of domain-restricted fan-out queries target the bare domain, only 9.1% a URL pathconfirmedEven sharper in our data: 86 domain parameters across 5 measurement days (20-24 Aug, deduplicated per run), none with a path. German prompt set, smaller sample - we do not see his 9.1% path share.
45% of domains appear only once, 18.1% form a hard core recurring near-dailyconfirmedSame pattern here: 25 unique domains in 5 days, 28% on one day only, 9 domains (36%) on 4-5 of 5 days - core: openai.com, reuters.com, tagesschau.de (all five days).
URL paths as targets are fleeting: 64.9% one-off, only 3.5% hard coreunverifiableIn our pipe calls the fourth parameter never carries a path; the URL level of his analysis does not occur in our data.
The domain is a stable, learned signal - the URL level is dynamic and re-decided per runconfirmedOur per-question stable domain lists against daily-changing result lists (grouped-webpages) show the same split.

Related

promptwatch-2026-08-10 ray-2026-08-17 mohanadasan-2026-08-21 konitzny-2026-06-22